$\mathrm{TRIZ}^{a}$: Guiding Agent Evolution from Pattern Recognition to Solution Invention

Wenyin Liu; Yiheng Huang; Kai Wang

cs.AIOct 3, 2026

$\mathrm{TRIZ}^{a}$ : Guiding Agent Evolution from Pattern Recognition to Solution Invention

Authors: Wenyin Liu, Yiheng Huang, Kai Wang

Organizations: Guangdong University of Technology · Beijing University of Posts and Telecommunications · Beijing Denglu Technology Ltd

Abstract

We propose $\mathrm{TRIZ}^{a}$ (TRIZ exponentiated by an agent), a general R&D automation paradigm that combines TRIZ inventive theory with LLM-driven agent evolutionary search. TRIZ's 40 inventive principles and contradiction matrix provide structured, explainable directions for solution generation, replacing random or untyped mutation with theory-guided ideation. Functional information (FI), operationalized under a frozen reference contract, is combined with TRIZ Ideality to measure useful and harmful function on a commensurable information scale, while hard gates keep promotion distinct from metric improvement. We validate $\mathrm{TRIZ}^{a}$ in cybersecurity--an adversarial and rapidly evolving domain--on PowerDuck GOOSE, CICIoT2023, and CIC-DDoS2019. Under paired-rerun protocols with protocol fingerprinting and hard-gate validation, the legacy experiments yield absolute F1 improvements of $+2.88$ , $+4.23$ , and $+0.15$ percentage points, respectively. A completed 45-activity CICIoT2023 campaign further increases macro-F1 from $0.8325$ to $0.8483$ , but does not pass its frozen promotion gate. Every result remains traceable from contradiction identification and TRIZ principle selection to code transformation, evaluation metrics, and promotion decision.

Figures & tables

Figure 1: System architecture of TRIZ a .

Principle	Typed transformation	Cybersecurity example
Segmentation (1)	Partition components or decisions behind a stable interface.	Per-family detectors with a shared output schema.
Local Quality (3)	Dispatch context-specific behavior under a registered context.	Class-aware scaling or decision boundaries.
Prior Action (10)	Prepare reusable state bef critical path.	Train-only feature cache bound to a split digest.
Parameter Change (35)	Change a type, threshold, loss, weight, or operating point.	Balanced class weights or calibrated loss.

Table 1: Representative TRIZ principles as executable transformation contracts.

Task / layer	Comparison	Observed evidence	Status and boundary
PowerDuck legacy	Paired weighted-F1	$0.9694\rightarrow 0.9982$ ( $+2.88$ pp)	complete ; no principle-level attribution
CICIoT legacy	Paired macro-F1	$0.8326\rightarrow 0.8749$ ( $+4.23$ pp)	downstream incumbent, not formal V2
CIC-DDoS legacy	Paired weighted-F1	$0.7470\rightarrow 0.7485$ ( $+0.15$ pp)	complete ; final ten-fold evaluation
CICIoT formal V2	45-activity campaign	macro-F1 $+0.0158$ ; recall $+0.018$ ; FPR $-0.0017$	complete evidence; not promoted
Campaign08 audit	2,040 controller decisions	balanced policy: promotion F1 $0.811$ , specificity $0.989$	descriptive nested aggregate
V17/V18 downstream	Multi-metric sequence	V17 retained for study; V18 deteriorated on three metrics	negative evidence preserved

Table 2: Evidence ladder. Results with different decision boundaries are not merged.

Figure 2: Experimental evidence at two decision levels. (A) Final same-protocol F1 outcomes, including the complete ten-fold CIC-DDoS evaluation and formal V2; the formal V2 candidate was not promoted. (B) Campaign08 policy trade-off. Safety-first V2 nearly maximizes specificity but loses recall and promotion F1; balanced V2 recovers useful decisions while retaining

0.989

specificity. Campaign08 is a descriptive aggregate of 2,040 nested decisions.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Artifact	Recorded content	Role
project.toml	claim ID, task, constraints, data and split IDs, protocol ID, and budgets	project and protocol record
triz_context.json	contradiction, improving and worsening parameters, principle set, and sources	contradiction record
triz_strategy_plan.md	declared transformations, arguments, prohibited changes, and required evidence	typed transformation plan
scores.jsonl	parent and candidate IDs, protocol ID, transformations, seeds, metrics, FI bounds, evidence levels, gates, and decisions	append-only evaluation record
SHA256SUMS.txt	byte digests for files included in the released artifact package	file identity and integrity check

Appendix

Table 3: Traceability record schema and planned archive-relative filenames.

No.	Principle	Typed transformation class
1	Segmentation	Partition modules, data, decisions, or search regions behind a stable interface.
2	Taking out	Isolate a confounder, harmful path, privileged component, or expensive stage.
3	Local quality	Dispatch context-specific policy under a registered context type.
4	Asymmetry	Apply different rules to unequal classes, risks, or resource directions.
5	Merging	Compose compatible stages, features, models, or evaluation passes.
6	Universality	Replace duplicated components with one multi-role typed service.

Appendix

Table 4: TRIZ principles and software/agent transformation classes.

Task	Primary metric	Useful event	Key harm/integrity checks
PowerDuck	weighted-F1, 15 classes	paired improvement under identical split	no embedding refit; class coverage; 10-shot boundary
CICIoT2023	macro-F1 across held-out families	family/seed improvement and recall	detector FPR, attack recall, split and seed identity
CIC-DDoS2019	weighted-F1 over folds	paired fold improvement	train-only feature fit, complete fold ledger
Campaign08	promotion F1 and balanced accuracy	correct promotion decision	false promotion, nested decision dependence

Appendix

Table 5: Registered task contracts.

Configuration/policy	Macro-F1	Recall	Precision	Specificity
RF100 formal baseline	0.8325	–	–	–
Formal V2 selected	0.8483	$+0.018$	–	FPR $-0.0017$
Original V1 controller	–	0.346	0.952	0.963
Safety-first V2 controller	–	0.257	0.997	0.998
Balanced V2 controller	–	0.686	0.993	0.989

Appendix

Table 6: Extended CICIoT2023 evidence. Promotion metrics describe the controller, not the detector.

Model	Macro-F1	MCC	Detector FPR	Attack recall
V1/RF100	0.8325	0.7153	0.0504	0.9052
V14	0.8749	0.7746	0.0228	0.8779
V17	0.8753	0.7755	0.0231	0.8809
V18	0.8752	0.7755	0.0232	0.8811

Appendix

Table 7: Downstream metric values shown in Figure 3 .

Figure 3: Downstream CICIoT2023 evolution. Relative to V1/RF100, V17 changes macro-F1 by

+5.13\%

, MCC by

+8.42\%

, and detector FPR by

-54.0\%

, while attack recall remains

2.43

pp below baseline. V18 is rejected. This path is separate from formal V2 selection.

Automated Algorithm Discovery

Automated algorithm discovery develops methods that generate and test executable algorithms for optimization, science, software, and machine learning.

See all 170 papers

CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly
May 25, 2026Yihe Fan, Changyi Li, Lichen Xu +4LLM Agent Self-Improvement Automated Penetration Testing
Fudan University, Shanghai, China · Shanghai Innovation Institute, Shanghai, China · Shanghai Pudong Research Institute of Cryptology
Adaptive Instruction Composition for Automated LLM Red-Teaming
Apr 22, 2026Jesse Zymet, Andy Luo, Swapnil Shinde +2Adversarial Prompt Generation LLM Red Teaming
Capital One, AI Foundations
RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution
Aug 27, 2026Junjie Zhang, Hui Liu, Kecheng Chen +3Automated Red Teaming LLM Agent Security
City University of Hong Kong · Shenzhen MSU-BIT University

May 25, 2026J/K move · Enter open · S save

CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly

Yihe Fan, Changyi Li, Lichen Xu, Xudong Pan+3

Fudan University, Shanghai, China · Shanghai Innovation Institute, Shanghai, China · Shanghai Pudong Research Institute of Cryptology

LLM Agent Self-Improvement Automated Penetration Testing LLMs for Cybersecurity Self-Evolving Agents

Open paper AskPDF arXiv

LLM-based agents are increasingly used for cybersecurity tasks, but most existing systems rely on fixed, human-designed scaffolds that struggle to adapt across diverse targets and failure modes. We introduce CyberEvolver, a self-evolving cybersecurity agent framework that iteratively revises its own scaffold based on experience from failed execution attempts. Self-evolution in cybersecurity is challenging because the space of possible scaffold changes is largely unstructured, execution feedback is sparse and often obscured by the environment, and low-diversity updates can cause errors to compound over repeated iterations. CyberEvolver addresses these challenges with a four-layer evolvable agent architecture that decomposes scaffold optimization into structured components, a trace-to-diagnosis mechanism that converts noisy execution logs into actionable revision signals, and a population-based beam search strategy that preserves diverse agent variants during evolution. We evaluate CyberEvolver on CTF challenges, vulnerability exploitation, and penetration-testing tasks using four open-source LLMs. Across these settings, CyberEvolver improves the seed agent's success rate by

13.6

,% on average, and outperforms six human-designed cybersecurity agents as well as two self-improvement methods adapted from other domains. These results suggest that scaffold self-evolution is a promising direction for building adaptive LLM agents for security testing.

Abstract

Automated Algorithm Discovery

Explore similar work

CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly

Adaptive Instruction Composition for Automated LLM Red-Teaming

RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution